The Agent Skills Directory

[SAFE]: The skill strictly defines read-only operations for reviewing code changes. It explicitly prohibits file mutation, quality loop execution, or git state modification.
[SAFE]: No evidence of hardcoded credentials, malicious obfuscation, or persistence mechanisms was found.
[INDIRECT_PROMPT_INJECTION]: The skill possesses a natural attack surface as it is designed to ingest and process untrusted external data, such as PR bodies and diffs. However, it implements risk-reduction measures:
Ingestion points: PRs, commits, local diffs, and worktree identity in SKILL.md.
Boundary markers: The skill requires the use of LeafContext and CandidateRef as authoritative continuation boundaries and adheres to a canonical Verification Record schema for output.
Capability inventory: Limited to read-only access of the provided file scope.
Sanitization: The instructions mandate inspecting raw files rather than trusting summaries and forbid the reconstruction of second-hand gate policies.

code-review